Search

As a new technology for model compression, self-knowledge distillation (SKD) avoids the large computational overhead of training the teacher model seen with traditional knowledge distillation. However, existing SKD methods pay attention to the knowledge transfer between deep and shallow layers of the network but ignore the mutual learning between shallow branches. This paper proposes a many-objective evolutionary self-knowledge distillation framework (MaOESKD) to guide the knowledge fusion between branches in the SKD neural network. This framework embeds an optimization module and a temporary branch into the multi-branch SKD network. The optimization module includes a many-objective adaptive weight optimization model (MaAWOM) and an many evolutionary optimization algorithm based on multi-strategy consensus mechanism (MaOEA-MCM); meanwhile, the temporary branch performs linear weighted fusion. In the MOBWOM, the weight of different branches in knowledge fusion is taken as the decision variable, and the mutual information, covariance, KL divergence between branch output features, and the total information of each branch are taken as the optimization objective. The MSCMEA integrates several state-of-the-art individual selection strategies in the field of evolutionary algorithms. It includes shift density estimation (SDE), penalized boundary intersection (PBI), balanced fitness estimation (BFE), and adaptive position transformation (APT). The accuracy of MOESKD achieves 99.70, 95.74 and 78.21 in MNIST, CIFAR-10 and CIFAR-100. [ABSTRACT FROM AUTHOR]

The application of semantic technologies, particularly ontologies, in the realm of multi-objective evolutionary algorithms is overlook despite their effectiveness in knowledge representation. In this paper, we introduce MOODY, an ontology specifically tailored to formalize these kinds of algorithms, encompassing their respective parameters, and multi-objective optimization problems based on a characterization of their search space landscapes. MOODY is designed to be particularly applicable in automatic algorithm configuration, which involves the search of the parameters of an optimization algorithm to optimize its performance. In this context, we observe a notable absence of standardized components, parameters, and related considerations, such as problem characteristics and algorithm configurations. This lack of standardization introduces difficulties in the selection of valid component combinations and in the re-use of algorithmic configurations between different algorithm implementations. MOODY offers a means to infuse semantic annotations into the configurations found by automatic tools, enabling efficient querying of the results and seamless integration across diverse sources through their incorporation into a knowledge graph. We validate our proposal by presenting four case studies. • A semantic model is proposed for consolidating multi-objective optimization knowledge. • A knowledge graph is presented, as well as the tools to include new RDF data. • The model allows for semantic reasoning on evolutionary algorithm configurations. • The proposal is validated on three use cases with data from experiments. • Moody enables knowledge-based recommendations of configurations for algorithms. [ABSTRACT FROM AUTHOR]

Dynamic multi-objective optimization problems (DMOPs) are multi-objective problems that are influenced by dynamically changing environmental parameters. Most current algorithms for solving DMOPs only respond to dynamic changes in the decision space or objective space and also ignore the impact of the type of DMOPs on the algorithm. The changes in the Pareto-optimal solution (POS) and Pareto-optimal front (POF) may affect the type of change in DMOPs. Therefore, this paper proposed an adaptive dynamic multi-objective evolutionary algorithm for type detection (TDA-DMOEA). First, the dynamic detection operator is designed to identify the types of dynamic problems. The Wilcoxon signed-rank test and Hyper Volume (HV) are used to detect the difference of POS and POF in two adjacent environments respectively. Then, different response strategies are designed to cope with different types of changes in DMOP. In particular, a multi-angle-based transfer learning method (MA-TL) with a closed kernel function is derived when faced with simultaneous changes in POS and POF. Finally, a comprehensive study of the commonly used benchmark set of DMOPs is presented, and the proposed algorithm achieves better performance in optimizing DMOPs. [ABSTRACT FROM AUTHOR]

Multi-objective optimization problems with interval parameter (IMOPs) are among the most critical optimization problems in practical applications. However, in contrast to deterministic multi-objective optimization, few studies have addressed IMOPs. Moreover, the uncertainty in such problems makes the convergence and diversity of the algorithm more challenging. Therefore, this paper proposes an interval multi-objective optimization algorithm based on elite genetic strategy (EG-IMOEA). First, existing interval dominance relations cannot comprehensively determine which of the two intervals is better when their midpoints are equal but their upper and lower limits are unequal. Therefore, a conditional-based interval confidence dominance relation is proposed that considers both the average level of convergence and value of the minimum lower limit of intervals. The interval crowding distance (ICD) applied to multiple objectives is then defined to evaluate the solutions more completely. Furthermore, an elite genetic strategy for crossover and mutation is proposed to generate better offspring. The proposed algorithm was evaluated on nine benchmark test problems and compared with the typical algorithm as well as state-of-the-art algorithms. The results showed that this algorithm outperformed the comparison algorithms in terms of convergence, diversity, imprecision, and uniform distribution. [ABSTRACT FROM AUTHOR]

The distribution characteristic of populations is one of the main characteristics of population. The pattern of distribution characteristics determines the probability law of individual values. At the same time, the initialization of the population lays a foundation for the iterative process of the swarm intelligence optimization algorithm. To reveal the influence of population distribution characteristics on the swarm intelligence optimization algorithm, this paper proposes a variety of search strategies based on the populations with different distribution characteristics and analyzes the influence of population distribution characteristics on optimization process by comparing the test results of the optimization algorithm after implementing these strategies. First, a new population generator is designed that can transform the same initial population into a population with uniform and central peaking distributions. On this basis, the two kinds of populations are applied to the global and local search stages of the optimization process, and four different search strategies are formed. Among them, the global and local search strategies based on a uniformly distributed population are the traditional methods. Finally, the performance of the optimization algorithm using different search strategies is evaluated through 29 test problems from the CEC2017 benchmark for 10, 30, 50 and 100 dimensions. In addition, the algorithms are applied to solve the TSP problem. For CEC2017 in 100D, only 13 of the 29 test functions achieve the best optimization effect by using the traditional method, while the other 16 test functions achieve better search results by using the other three search strategies. The analysis shows that the population distribution characteristics have a great influence on the population optimization algorithm. The performance of the algorithm with different population distribution combination strategies is statistically superior to the traditional algorithm with a uniform distribution population, as revealed in the test functions ranging from 38.7% to 62.9% for different dimensions. By reconstructing populations with different distribution characteristics, the overall efficiency of the swarm intelligence optimization algorithm can be greatly improved. [ABSTRACT FROM AUTHOR]

In this paper, a feature selection (FS) method is proposed to identify key quality features (KQFs) in complex manufacturing processes. We propose a multi-objective binary particle swarm optimization algorithm, called MPBPSO, with three new components to optimize a bi-objective FS model of maximizing the geometric mean (GM) measure and minimizing the number of selected features. First, MPBPSO uses a modified probability-based solution update (PSU) mechanism which utilizes a flipping vector to update particles. A mutation operator with three basic operations, i.e., add, eliminate, and interchange, is also utilized in MPBPSO to improve the exploration performance. Second, a strategy combining the Pareto dominance concept with a distance measure is proposed for MPBPSO to update p b e s t (personal best position). Finally, a selection strategy based on the roulette wheel selection is proposed to determine the g b e s t (global best position) from the non-dominated set during iterations. The experimental results on four datasets have shown that the proposed FS method can identify a small number of KQFs that have good predictive ability for product quality. Further analysis indicates that MPBPSO obtains better search performance than eight benchmark optimization algorithms and the new components in MPBPSO are effective for improving its search performance. • Quality feature selection is addressed by a multi-objective optimization model. • A multi-objective binary PSO algorithm is proposed for the model. • Probability-based mechanism is combined with mutation for updating solutions. • New pbest update and gbest selection strategies are proposed. • Experiments verify the effectiveness of the model and optimization algorithm. [ABSTRACT FROM AUTHOR]